codegen: fix compute_va_type issue with Type (#46953)

Issue noted in #42498. This should be the same as
Core.Compiler.tuple_tfunc. Otherwise we might accidentally constant-fold
something like:

     code_llvm((x...) -> x isa Tuple{Type{Tuple{Any}},Int}, (Type{Tuple{Any}}, Int))

to return true. This is rarely a compile-sig in practice, so it does not
usually affect code, but is observable there in the IR.

src/codegen.cpp

@@ -2144,12 +2144,7 @@ static void jl_init_function(Function *F)
 
 static std::pair<bool, bool> uses_specsig(jl_method_instance_t *lam, jl_value_t *rettype, bool prefer_specsig)
 {
-    size_t nreq = jl_is_method(lam->def.method) ? lam->def.method->nargs : 0;
-    int va = 0;
-    if (nreq > 0 && lam->def.method->isva) {
-        nreq--;
-        va = 1;
-    }
+    int va = lam->def.method->isva;
     jl_value_t *sig = lam->specTypes;
     bool needsparams = false;
     if (jl_is_method(lam->def.method)) {
@@ -6594,13 +6589,20 @@ get_specsig_di(jl_codectx_t &ctx, jl_debugcache_t &debuginfo, jl_value_t *rt, jl
     return dbuilder.createSubroutineType(dbuilder.getOrCreateTypeArray(ditypes));
 }
 
+/* aka Core.Compiler.tuple_tfunc */
 static jl_datatype_t *compute_va_type(jl_method_instance_t *lam, size_t nreq)
 {
     size_t nvargs = jl_nparams(lam->specTypes)-nreq;
     jl_svec_t *tupargs = jl_alloc_svec(nvargs);
     JL_GC_PUSH1(&tupargs);
     for (size_t i = nreq; i < jl_nparams(lam->specTypes); ++i) {
         jl_value_t *argType = jl_nth_slot_type(lam->specTypes, i);
+        if (is_uniquerep_Type(argType))
+            argType = jl_typeof(jl_tparam0(argType));
+        else if (jl_has_intersect_type_not_kind(argType)) {
+            jl_value_t *ts[2] = {argType, (jl_value_t*)jl_type_type};
+            argType = jl_type_union(ts, 2);
+        }
         jl_svecset(tupargs, i-nreq, argType);
     }
     jl_datatype_t *typ = jl_apply_tuple_type(tupargs);